aacenc: Template quantize_and_encode_band_cost().
[ffmpeg.git] / libavcodec / aaccoder.c
1 /*
2  * AAC coefficients encoder
3  * Copyright (C) 2008-2009 Konstantin Shishkov
4  *
5  * This file is part of FFmpeg.
6  *
7  * FFmpeg is free software; you can redistribute it and/or
8  * modify it under the terms of the GNU Lesser General Public
9  * License as published by the Free Software Foundation; either
10  * version 2.1 of the License, or (at your option) any later version.
11  *
12  * FFmpeg is distributed in the hope that it will be useful,
13  * but WITHOUT ANY WARRANTY; without even the implied warranty of
14  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
15  * Lesser General Public License for more details.
16  *
17  * You should have received a copy of the GNU Lesser General Public
18  * License along with FFmpeg; if not, write to the Free Software
19  * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
20  */
21
22 /**
23  * @file
24  * AAC coefficients encoder
25  */
26
27 /***********************************
28  *              TODOs:
29  * speedup quantizer selection
30  * add sane pulse detection
31  ***********************************/
32
33 #include <float.h>
34 #include "avcodec.h"
35 #include "put_bits.h"
36 #include "aac.h"
37 #include "aacenc.h"
38 #include "aactab.h"
39
40 /** bits needed to code codebook run value for long windows */
41 static const uint8_t run_value_bits_long[64] = {
42      5,  5,  5,  5,  5,  5,  5,  5,  5,  5,  5,  5,  5,  5,  5,  5,
43      5,  5,  5,  5,  5,  5,  5,  5,  5,  5,  5,  5,  5,  5,  5, 10,
44     10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10,
45     10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 15
46 };
47
48 /** bits needed to code codebook run value for short windows */
49 static const uint8_t run_value_bits_short[16] = {
50     3, 3, 3, 3, 3, 3, 3, 6, 6, 6, 6, 6, 6, 6, 6, 9
51 };
52
53 static const uint8_t *run_value_bits[2] = {
54     run_value_bits_long, run_value_bits_short
55 };
56
57
58 /**
59  * Quantize one coefficient.
60  * @return absolute value of the quantized coefficient
61  * @see 3GPP TS26.403 5.6.2 "Scalefactor determination"
62  */
63 static av_always_inline int quant(float coef, const float Q)
64 {
65     float a = coef * Q;
66     return sqrtf(a * sqrtf(a)) + 0.4054;
67 }
68
69 static void quantize_bands(int *out, const float *in, const float *scaled,
70                            int size, float Q34, int is_signed, int maxval)
71 {
72     int i;
73     double qc;
74     for (i = 0; i < size; i++) {
75         qc = scaled[i] * Q34;
76         out[i] = (int)FFMIN(qc + 0.4054, (double)maxval);
77         if (is_signed && in[i] < 0.0f) {
78             out[i] = -out[i];
79         }
80     }
81 }
82
83 static void abs_pow34_v(float *out, const float *in, const int size)
84 {
85 #ifndef USE_REALLY_FULL_SEARCH
86     int i;
87     for (i = 0; i < size; i++) {
88         float a = fabsf(in[i]);
89         out[i] = sqrtf(a * sqrtf(a));
90     }
91 #endif /* USE_REALLY_FULL_SEARCH */
92 }
93
94 static const uint8_t aac_cb_range [12] = {0, 3, 3, 3, 3, 9, 9, 8, 8, 13, 13, 17};
95 static const uint8_t aac_cb_maxval[12] = {0, 1, 1, 2, 2, 4, 4, 7, 7, 12, 12, 16};
96
97 /**
98  * Calculate rate distortion cost for quantizing with given codebook
99  *
100  * @return quantization distortion
101  */
102 static av_always_inline float quantize_and_encode_band_cost_template(
103                                 struct AACEncContext *s,
104                                 PutBitContext *pb, const float *in,
105                                 const float *scaled, int size, int scale_idx,
106                                 int cb, const float lambda, const float uplim,
107                                 int *bits, int BT_ZERO, int BT_UNSIGNED,
108                                 int BT_PAIR, int BT_ESC)
109 {
110     const float IQ = ff_aac_pow2sf_tab[200 + scale_idx - SCALE_ONE_POS + SCALE_DIV_512];
111     const float  Q = ff_aac_pow2sf_tab[200 - scale_idx + SCALE_ONE_POS - SCALE_DIV_512];
112     const float CLIPPED_ESCAPE = 165140.0f*IQ;
113     int i, j, k;
114     float cost = 0;
115     const int dim = BT_PAIR ? 2 : 4;
116     int resbits = 0;
117     const float  Q34 = sqrtf(Q * sqrtf(Q));
118     const int range  = aac_cb_range[cb];
119     const int maxval = aac_cb_maxval[cb];
120     int off;
121
122     if (BT_ZERO) {
123         for (i = 0; i < size; i++)
124             cost += in[i]*in[i];
125         if (bits)
126             *bits = 0;
127         return cost * lambda;
128     }
129     if (!scaled) {
130         abs_pow34_v(s->scoefs, in, size);
131         scaled = s->scoefs;
132     }
133     quantize_bands(s->qcoefs, in, scaled, size, Q34, !BT_UNSIGNED, maxval);
134     if (BT_UNSIGNED) {
135         off = 0;
136     } else {
137         off = maxval;
138     }
139     for (i = 0; i < size; i += dim) {
140         const float *vec;
141         int *quants = s->qcoefs + i;
142         int curidx = 0;
143         int curbits;
144         float rd = 0.0f;
145         for (j = 0; j < dim; j++) {
146             curidx *= range;
147             curidx += quants[j] + off;
148         }
149             curbits =  ff_aac_spectral_bits[cb-1][curidx];
150             vec     = &ff_aac_codebook_vectors[cb-1][curidx*dim];
151             if (BT_UNSIGNED) {
152                 for (k = 0; k < dim; k++) {
153                     float t = fabsf(in[i+k]);
154                     float di;
155                     if (BT_ESC && vec[k] == 64.0f) { //FIXME: slow
156                         if (t >= CLIPPED_ESCAPE) {
157                             di = t - CLIPPED_ESCAPE;
158                             curbits += 21;
159                         } else {
160                             int c = av_clip(quant(t, Q), 0, 8191);
161                             di = t - c*cbrtf(c)*IQ;
162                             curbits += av_log2(c)*2 - 4 + 1;
163                         }
164                     } else {
165                         di = t - vec[k]*IQ;
166                     }
167                     if (vec[k] != 0.0f)
168                         curbits++;
169                     rd += di*di;
170                 }
171             } else {
172                 for (k = 0; k < dim; k++) {
173                     float di = in[i+k] - vec[k]*IQ;
174                     rd += di*di;
175                 }
176             }
177         cost    += rd * lambda + curbits;
178         resbits += curbits;
179         if (cost >= uplim)
180             return uplim;
181         if (pb) {
182         put_bits(pb, ff_aac_spectral_bits[cb-1][curidx], ff_aac_spectral_codes[cb-1][curidx]);
183         if (BT_UNSIGNED)
184             for (j = 0; j < dim; j++)
185                 if (ff_aac_codebook_vectors[cb-1][curidx*dim+j] != 0.0f)
186                     put_bits(pb, 1, in[i+j] < 0.0f);
187         if (BT_ESC) {
188             for (j = 0; j < 2; j++) {
189                 if (ff_aac_codebook_vectors[cb-1][curidx*2+j] == 64.0f) {
190                     int coef = av_clip(quant(fabsf(in[i+j]), Q), 0, 8191);
191                     int len = av_log2(coef);
192
193                     put_bits(pb, len - 4 + 1, (1 << (len - 4 + 1)) - 2);
194                     put_bits(pb, len, coef & ((1 << len) - 1));
195                 }
196             }
197         }
198         }
199     }
200
201     if (bits)
202         *bits = resbits;
203     return cost;
204 }
205
206 #define QUANTIZE_AND_ENCODE_BAND_COST_FUNC(NAME, BT_ZERO, BT_UNSIGNED, BT_PAIR, BT_ESC) \
207 static float quantize_and_encode_band_cost_ ## NAME(                                        \
208                                 struct AACEncContext *s,                                \
209                                 PutBitContext *pb, const float *in,                     \
210                                 const float *scaled, int size, int scale_idx,           \
211                                 int cb, const float lambda, const float uplim,          \
212                                 int *bits) {                                            \
213     return quantize_and_encode_band_cost_template(                                      \
214                                 s, pb, in, scaled, size, scale_idx,                     \
215                                 BT_ESC ? ESC_BT : cb, lambda, uplim, bits,              \
216                                 BT_ZERO, BT_UNSIGNED, BT_PAIR, BT_ESC);                 \
217 }
218
219 QUANTIZE_AND_ENCODE_BAND_COST_FUNC(ZERO,  1, 0, 0, 0)
220 QUANTIZE_AND_ENCODE_BAND_COST_FUNC(SQUAD, 0, 0, 0, 0)
221 QUANTIZE_AND_ENCODE_BAND_COST_FUNC(UQUAD, 0, 1, 0, 0)
222 QUANTIZE_AND_ENCODE_BAND_COST_FUNC(SPAIR, 0, 0, 1, 0)
223 QUANTIZE_AND_ENCODE_BAND_COST_FUNC(UPAIR, 0, 1, 1, 0)
224 QUANTIZE_AND_ENCODE_BAND_COST_FUNC(ESC,   0, 1, 1, 1)
225
226 static float (*quantize_and_encode_band_cost_arr[])(
227                                 struct AACEncContext *s,
228                                 PutBitContext *pb, const float *in,
229                                 const float *scaled, int size, int scale_idx,
230                                 int cb, const float lambda, const float uplim,
231                                 int *bits) = {
232     quantize_and_encode_band_cost_ZERO,
233     quantize_and_encode_band_cost_SQUAD,
234     quantize_and_encode_band_cost_SQUAD,
235     quantize_and_encode_band_cost_UQUAD,
236     quantize_and_encode_band_cost_UQUAD,
237     quantize_and_encode_band_cost_SPAIR,
238     quantize_and_encode_band_cost_SPAIR,
239     quantize_and_encode_band_cost_UPAIR,
240     quantize_and_encode_band_cost_UPAIR,
241     quantize_and_encode_band_cost_UPAIR,
242     quantize_and_encode_band_cost_UPAIR,
243     quantize_and_encode_band_cost_ESC,
244 };
245
246 #define quantize_and_encode_band_cost(                                  \
247                                 s, pb, in, scaled, size, scale_idx, cb, \
248                                 lambda, uplim, bits)                    \
249     quantize_and_encode_band_cost_arr[cb](                              \
250                                 s, pb, in, scaled, size, scale_idx, cb, \
251                                 lambda, uplim, bits)
252
253 static float quantize_band_cost(struct AACEncContext *s, const float *in,
254                                 const float *scaled, int size, int scale_idx,
255                                 int cb, const float lambda, const float uplim,
256                                 int *bits)
257 {
258     return quantize_and_encode_band_cost(s, NULL, in, scaled, size, scale_idx,
259                                          cb, lambda, uplim, bits);
260 }
261
262 static void quantize_and_encode_band(struct AACEncContext *s, PutBitContext *pb,
263                                      const float *in, int size, int scale_idx,
264                                      int cb, const float lambda)
265 {
266     quantize_and_encode_band_cost(s, pb, in, NULL, size, scale_idx, cb, lambda,
267                                   INFINITY, NULL);
268 }
269
270 static float find_max_val(int group_len, int swb_size, const float *scaled) {
271     float maxval = 0.0f;
272     int w2, i;
273     for (w2 = 0; w2 < group_len; w2++) {
274         for (i = 0; i < swb_size; i++) {
275             maxval = FFMAX(maxval, scaled[w2*128+i]);
276         }
277     }
278     return maxval;
279 }
280
281 static int find_min_book(float maxval, int sf) {
282     float Q = ff_aac_pow2sf_tab[200 - sf + SCALE_ONE_POS - SCALE_DIV_512];
283     float Q34 = sqrtf(Q * sqrtf(Q));
284     int qmaxval, cb;
285     qmaxval = maxval * Q34 + 0.4054f;
286     if      (qmaxval ==  0) cb = 0;
287     else if (qmaxval ==  1) cb = 1;
288     else if (qmaxval ==  2) cb = 3;
289     else if (qmaxval <=  4) cb = 5;
290     else if (qmaxval <=  7) cb = 7;
291     else if (qmaxval <= 12) cb = 9;
292     else                    cb = 11;
293     return cb;
294 }
295
296 /**
297  * structure used in optimal codebook search
298  */
299 typedef struct BandCodingPath {
300     int prev_idx; ///< pointer to the previous path point
301     float cost;   ///< path cost
302     int run;
303 } BandCodingPath;
304
305 /**
306  * Encode band info for single window group bands.
307  */
308 static void encode_window_bands_info(AACEncContext *s, SingleChannelElement *sce,
309                                      int win, int group_len, const float lambda)
310 {
311     BandCodingPath path[120][12];
312     int w, swb, cb, start, start2, size;
313     int i, j;
314     const int max_sfb  = sce->ics.max_sfb;
315     const int run_bits = sce->ics.num_windows == 1 ? 5 : 3;
316     const int run_esc  = (1 << run_bits) - 1;
317     int idx, ppos, count;
318     int stackrun[120], stackcb[120], stack_len;
319     float next_minrd = INFINITY;
320     int next_mincb = 0;
321
322     abs_pow34_v(s->scoefs, sce->coeffs, 1024);
323     start = win*128;
324     for (cb = 0; cb < 12; cb++) {
325         path[0][cb].cost     = 0.0f;
326         path[0][cb].prev_idx = -1;
327         path[0][cb].run      = 0;
328     }
329     for (swb = 0; swb < max_sfb; swb++) {
330         start2 = start;
331         size = sce->ics.swb_sizes[swb];
332         if (sce->zeroes[win*16 + swb]) {
333             for (cb = 0; cb < 12; cb++) {
334                 path[swb+1][cb].prev_idx = cb;
335                 path[swb+1][cb].cost     = path[swb][cb].cost;
336                 path[swb+1][cb].run      = path[swb][cb].run + 1;
337             }
338         } else {
339             float minrd = next_minrd;
340             int mincb = next_mincb;
341             next_minrd = INFINITY;
342             next_mincb = 0;
343             for (cb = 0; cb < 12; cb++) {
344                 float cost_stay_here, cost_get_here;
345                 float rd = 0.0f;
346                 for (w = 0; w < group_len; w++) {
347                     FFPsyBand *band = &s->psy.psy_bands[s->cur_channel*PSY_MAX_BANDS+(win+w)*16+swb];
348                     rd += quantize_band_cost(s, sce->coeffs + start + w*128,
349                                              s->scoefs + start + w*128, size,
350                                              sce->sf_idx[(win+w)*16+swb], cb,
351                                              lambda / band->threshold, INFINITY, NULL);
352                 }
353                 cost_stay_here = path[swb][cb].cost + rd;
354                 cost_get_here  = minrd              + rd + run_bits + 4;
355                 if (   run_value_bits[sce->ics.num_windows == 8][path[swb][cb].run]
356                     != run_value_bits[sce->ics.num_windows == 8][path[swb][cb].run+1])
357                     cost_stay_here += run_bits;
358                 if (cost_get_here < cost_stay_here) {
359                     path[swb+1][cb].prev_idx = mincb;
360                     path[swb+1][cb].cost     = cost_get_here;
361                     path[swb+1][cb].run      = 1;
362                 } else {
363                     path[swb+1][cb].prev_idx = cb;
364                     path[swb+1][cb].cost     = cost_stay_here;
365                     path[swb+1][cb].run      = path[swb][cb].run + 1;
366                 }
367                 if (path[swb+1][cb].cost < next_minrd) {
368                     next_minrd = path[swb+1][cb].cost;
369                     next_mincb = cb;
370                 }
371             }
372         }
373         start += sce->ics.swb_sizes[swb];
374     }
375
376     //convert resulting path from backward-linked list
377     stack_len = 0;
378     idx       = 0;
379     for (cb = 1; cb < 12; cb++)
380         if (path[max_sfb][cb].cost < path[max_sfb][idx].cost)
381             idx = cb;
382     ppos = max_sfb;
383     while (ppos > 0) {
384         cb = idx;
385         stackrun[stack_len] = path[ppos][cb].run;
386         stackcb [stack_len] = cb;
387         idx = path[ppos-path[ppos][cb].run+1][cb].prev_idx;
388         ppos -= path[ppos][cb].run;
389         stack_len++;
390     }
391     //perform actual band info encoding
392     start = 0;
393     for (i = stack_len - 1; i >= 0; i--) {
394         put_bits(&s->pb, 4, stackcb[i]);
395         count = stackrun[i];
396         memset(sce->zeroes + win*16 + start, !stackcb[i], count);
397         //XXX: memset when band_type is also uint8_t
398         for (j = 0; j < count; j++) {
399             sce->band_type[win*16 + start] =  stackcb[i];
400             start++;
401         }
402         while (count >= run_esc) {
403             put_bits(&s->pb, run_bits, run_esc);
404             count -= run_esc;
405         }
406         put_bits(&s->pb, run_bits, count);
407     }
408 }
409
410 static void codebook_trellis_rate(AACEncContext *s, SingleChannelElement *sce,
411                                   int win, int group_len, const float lambda)
412 {
413     BandCodingPath path[120][12];
414     int w, swb, cb, start, start2, size;
415     int i, j;
416     const int max_sfb  = sce->ics.max_sfb;
417     const int run_bits = sce->ics.num_windows == 1 ? 5 : 3;
418     const int run_esc  = (1 << run_bits) - 1;
419     int idx, ppos, count;
420     int stackrun[120], stackcb[120], stack_len;
421     float next_minrd = INFINITY;
422     int next_mincb = 0;
423
424     abs_pow34_v(s->scoefs, sce->coeffs, 1024);
425     start = win*128;
426     for (cb = 0; cb < 12; cb++) {
427         path[0][cb].cost     = run_bits+4;
428         path[0][cb].prev_idx = -1;
429         path[0][cb].run      = 0;
430     }
431     for (swb = 0; swb < max_sfb; swb++) {
432         start2 = start;
433         size = sce->ics.swb_sizes[swb];
434         if (sce->zeroes[win*16 + swb]) {
435             for (cb = 0; cb < 12; cb++) {
436                 path[swb+1][cb].prev_idx = cb;
437                 path[swb+1][cb].cost     = path[swb][cb].cost;
438                 path[swb+1][cb].run      = path[swb][cb].run + 1;
439             }
440         } else {
441             float minrd = next_minrd;
442             int mincb = next_mincb;
443             int startcb = sce->band_type[win*16+swb];
444             next_minrd = INFINITY;
445             next_mincb = 0;
446             for (cb = 0; cb < startcb; cb++) {
447                 path[swb+1][cb].cost = 61450;
448                 path[swb+1][cb].prev_idx = -1;
449                 path[swb+1][cb].run = 0;
450             }
451             for (cb = startcb; cb < 12; cb++) {
452                 float cost_stay_here, cost_get_here;
453                 float rd = 0.0f;
454                 for (w = 0; w < group_len; w++) {
455                     rd += quantize_band_cost(s, sce->coeffs + start + w*128,
456                                              s->scoefs + start + w*128, size,
457                                              sce->sf_idx[(win+w)*16+swb], cb,
458                                              0, INFINITY, NULL);
459                 }
460                 cost_stay_here = path[swb][cb].cost + rd;
461                 cost_get_here  = minrd              + rd + run_bits + 4;
462                 if (   run_value_bits[sce->ics.num_windows == 8][path[swb][cb].run]
463                     != run_value_bits[sce->ics.num_windows == 8][path[swb][cb].run+1])
464                     cost_stay_here += run_bits;
465                 if (cost_get_here < cost_stay_here) {
466                     path[swb+1][cb].prev_idx = mincb;
467                     path[swb+1][cb].cost     = cost_get_here;
468                     path[swb+1][cb].run      = 1;
469                 } else {
470                     path[swb+1][cb].prev_idx = cb;
471                     path[swb+1][cb].cost     = cost_stay_here;
472                     path[swb+1][cb].run      = path[swb][cb].run + 1;
473                 }
474                 if (path[swb+1][cb].cost < next_minrd) {
475                     next_minrd = path[swb+1][cb].cost;
476                     next_mincb = cb;
477                 }
478             }
479         }
480         start += sce->ics.swb_sizes[swb];
481     }
482
483     //convert resulting path from backward-linked list
484     stack_len = 0;
485     idx       = 0;
486     for (cb = 1; cb < 12; cb++)
487         if (path[max_sfb][cb].cost < path[max_sfb][idx].cost)
488             idx = cb;
489     ppos = max_sfb;
490     while (ppos > 0) {
491         if (idx < 0) abort();
492         cb = idx;
493         stackrun[stack_len] = path[ppos][cb].run;
494         stackcb [stack_len] = cb;
495         idx = path[ppos-path[ppos][cb].run+1][cb].prev_idx;
496         ppos -= path[ppos][cb].run;
497         stack_len++;
498     }
499     //perform actual band info encoding
500     start = 0;
501     for (i = stack_len - 1; i >= 0; i--) {
502         put_bits(&s->pb, 4, stackcb[i]);
503         count = stackrun[i];
504         memset(sce->zeroes + win*16 + start, !stackcb[i], count);
505         //XXX: memset when band_type is also uint8_t
506         for (j = 0; j < count; j++) {
507             sce->band_type[win*16 + start] =  stackcb[i];
508             start++;
509         }
510         while (count >= run_esc) {
511             put_bits(&s->pb, run_bits, run_esc);
512             count -= run_esc;
513         }
514         put_bits(&s->pb, run_bits, count);
515     }
516 }
517
518 typedef struct TrellisPath {
519     float cost;
520     int prev;
521 } TrellisPath;
522
523 #define TRELLIS_STAGES 121
524 #define TRELLIS_STATES (SCALE_MAX_DIFF+1)
525
526 static void search_for_quantizers_anmr(AVCodecContext *avctx, AACEncContext *s,
527                                        SingleChannelElement *sce,
528                                        const float lambda)
529 {
530     int q, w, w2, g, start = 0;
531     int i, j;
532     int idx;
533     TrellisPath paths[TRELLIS_STAGES][TRELLIS_STATES];
534     int bandaddr[TRELLIS_STAGES];
535     int minq;
536     float mincost;
537     float q0f = FLT_MAX, q1f = 0.0f, qnrgf = 0.0f;
538     int q0, q1, qcnt = 0;
539
540     for (i = 0; i < 1024; i++) {
541         float t = fabsf(sce->coeffs[i]);
542         if (t > 0.0f) {
543             q0f = FFMIN(q0f, t);
544             q1f = FFMAX(q1f, t);
545             qnrgf += t*t;
546             qcnt++;
547         }
548     }
549
550     if (!qcnt) {
551         memset(sce->sf_idx, 0, sizeof(sce->sf_idx));
552         memset(sce->zeroes, 1, sizeof(sce->zeroes));
553         return;
554     }
555
556     //minimum scalefactor index is when minimum nonzero coefficient after quantizing is not clipped
557     q0 = av_clip_uint8(log2(q0f)*4 - 69 + SCALE_ONE_POS - SCALE_DIV_512);
558     //maximum scalefactor index is when maximum coefficient after quantizing is still not zero
559     q1 = av_clip_uint8(log2(q1f)*4 +  6 + SCALE_ONE_POS - SCALE_DIV_512);
560     //av_log(NULL, AV_LOG_ERROR, "q0 %d, q1 %d\n", q0, q1);
561     if (q1 - q0 > 60) {
562         int q0low  = q0;
563         int q1high = q1;
564         //minimum scalefactor index is when maximum nonzero coefficient after quantizing is not clipped
565         int qnrg = av_clip_uint8(log2(sqrt(qnrgf/qcnt))*4 - 31 + SCALE_ONE_POS - SCALE_DIV_512);
566         q1 = qnrg + 30;
567         q0 = qnrg - 30;
568     //av_log(NULL, AV_LOG_ERROR, "q0 %d, q1 %d\n", q0, q1);
569         if (q0 < q0low) {
570             q1 += q0low - q0;
571             q0  = q0low;
572         } else if (q1 > q1high) {
573             q0 -= q1 - q1high;
574             q1  = q1high;
575         }
576     }
577     //av_log(NULL, AV_LOG_ERROR, "q0 %d, q1 %d\n", q0, q1);
578
579     for (i = 0; i < TRELLIS_STATES; i++) {
580         paths[0][i].cost    = 0.0f;
581         paths[0][i].prev    = -1;
582     }
583     for (j = 1; j < TRELLIS_STAGES; j++) {
584         for (i = 0; i < TRELLIS_STATES; i++) {
585             paths[j][i].cost    = INFINITY;
586             paths[j][i].prev    = -2;
587         }
588     }
589     idx = 1;
590     abs_pow34_v(s->scoefs, sce->coeffs, 1024);
591     for (w = 0; w < sce->ics.num_windows; w += sce->ics.group_len[w]) {
592         start = w*128;
593         for (g = 0; g < sce->ics.num_swb; g++) {
594             const float *coefs = sce->coeffs + start;
595             float qmin, qmax;
596             int nz = 0;
597
598             bandaddr[idx] = w * 16 + g;
599             qmin = INT_MAX;
600             qmax = 0.0f;
601             for (w2 = 0; w2 < sce->ics.group_len[w]; w2++) {
602                 FFPsyBand *band = &s->psy.psy_bands[s->cur_channel*PSY_MAX_BANDS+(w+w2)*16+g];
603                 if (band->energy <= band->threshold || band->threshold == 0.0f) {
604                     sce->zeroes[(w+w2)*16+g] = 1;
605                     continue;
606                 }
607                 sce->zeroes[(w+w2)*16+g] = 0;
608                 nz = 1;
609                 for (i = 0; i < sce->ics.swb_sizes[g]; i++) {
610                     float t = fabsf(coefs[w2*128+i]);
611                     if (t > 0.0f)
612                         qmin = FFMIN(qmin, t);
613                     qmax = FFMAX(qmax, t);
614                 }
615             }
616             if (nz) {
617                 int minscale, maxscale;
618                 float minrd = INFINITY;
619                 float maxval;
620                 //minimum scalefactor index is when minimum nonzero coefficient after quantizing is not clipped
621                 minscale = av_clip_uint8(log2(qmin)*4 - 69 + SCALE_ONE_POS - SCALE_DIV_512);
622                 //maximum scalefactor index is when maximum coefficient after quantizing is still not zero
623                 maxscale = av_clip_uint8(log2(qmax)*4 +  6 + SCALE_ONE_POS - SCALE_DIV_512);
624                 minscale = av_clip(minscale - q0, 0, TRELLIS_STATES - 1);
625                 maxscale = av_clip(maxscale - q0, 0, TRELLIS_STATES);
626                 maxval = find_max_val(sce->ics.group_len[w], sce->ics.swb_sizes[g], s->scoefs+start);
627                 for (q = minscale; q < maxscale; q++) {
628                     float dist = 0;
629                     int cb = find_min_book(maxval, sce->sf_idx[w*16+g]);
630                     for (w2 = 0; w2 < sce->ics.group_len[w]; w2++) {
631                         FFPsyBand *band = &s->psy.psy_bands[s->cur_channel*PSY_MAX_BANDS+(w+w2)*16+g];
632                         dist += quantize_band_cost(s, coefs + w2*128, s->scoefs + start + w2*128, sce->ics.swb_sizes[g],
633                                                    q + q0, cb, lambda / band->threshold, INFINITY, NULL);
634                     }
635                     minrd = FFMIN(minrd, dist);
636
637                     for (i = 0; i < q1 - q0; i++) {
638                         float cost;
639                         cost = paths[idx - 1][i].cost + dist
640                                + ff_aac_scalefactor_bits[q - i + SCALE_DIFF_ZERO];
641                         if (cost < paths[idx][q].cost) {
642                             paths[idx][q].cost    = cost;
643                             paths[idx][q].prev    = i;
644                         }
645                     }
646                 }
647             } else {
648                 for (q = 0; q < q1 - q0; q++) {
649                     paths[idx][q].cost = paths[idx - 1][q].cost + 1;
650                     paths[idx][q].prev = q;
651                 }
652             }
653             sce->zeroes[w*16+g] = !nz;
654             start += sce->ics.swb_sizes[g];
655             idx++;
656         }
657     }
658     idx--;
659     mincost = paths[idx][0].cost;
660     minq    = 0;
661     for (i = 1; i < TRELLIS_STATES; i++) {
662         if (paths[idx][i].cost < mincost) {
663             mincost = paths[idx][i].cost;
664             minq = i;
665         }
666     }
667     while (idx) {
668         sce->sf_idx[bandaddr[idx]] = minq + q0;
669         minq = paths[idx][minq].prev;
670         idx--;
671     }
672     //set the same quantizers inside window groups
673     for (w = 0; w < sce->ics.num_windows; w += sce->ics.group_len[w])
674         for (g = 0;  g < sce->ics.num_swb; g++)
675             for (w2 = 1; w2 < sce->ics.group_len[w]; w2++)
676                 sce->sf_idx[(w+w2)*16+g] = sce->sf_idx[w*16+g];
677 }
678
679 /**
680  * two-loop quantizers search taken from ISO 13818-7 Appendix C
681  */
682 static void search_for_quantizers_twoloop(AVCodecContext *avctx,
683                                           AACEncContext *s,
684                                           SingleChannelElement *sce,
685                                           const float lambda)
686 {
687     int start = 0, i, w, w2, g;
688     int destbits = avctx->bit_rate * 1024.0 / avctx->sample_rate / avctx->channels;
689     float dists[128], uplims[128];
690     int fflag, minscaler;
691     int its  = 0;
692     int allz = 0;
693     float minthr = INFINITY;
694
695     //XXX: some heuristic to determine initial quantizers will reduce search time
696     memset(dists, 0, sizeof(dists));
697     //determine zero bands and upper limits
698     for (w = 0; w < sce->ics.num_windows; w += sce->ics.group_len[w]) {
699         for (g = 0;  g < sce->ics.num_swb; g++) {
700             int nz = 0;
701             float uplim = 0.0f;
702             for (w2 = 0; w2 < sce->ics.group_len[w]; w2++) {
703                 FFPsyBand *band = &s->psy.psy_bands[s->cur_channel*PSY_MAX_BANDS+(w+w2)*16+g];
704                 uplim += band->threshold;
705                 if (band->energy <= band->threshold || band->threshold == 0.0f) {
706                     sce->zeroes[(w+w2)*16+g] = 1;
707                     continue;
708                 }
709                 nz = 1;
710             }
711             uplims[w*16+g] = uplim *512;
712             sce->zeroes[w*16+g] = !nz;
713             if (nz)
714                 minthr = FFMIN(minthr, uplim);
715             allz = FFMAX(allz, nz);
716         }
717     }
718     for (w = 0; w < sce->ics.num_windows; w += sce->ics.group_len[w]) {
719         for (g = 0;  g < sce->ics.num_swb; g++) {
720             if (sce->zeroes[w*16+g]) {
721                 sce->sf_idx[w*16+g] = SCALE_ONE_POS;
722                 continue;
723             }
724             sce->sf_idx[w*16+g] = SCALE_ONE_POS + FFMIN(log2(uplims[w*16+g]/minthr)*4,59);
725         }
726     }
727
728     if (!allz)
729         return;
730     abs_pow34_v(s->scoefs, sce->coeffs, 1024);
731     //perform two-loop search
732     //outer loop - improve quality
733     do {
734         int tbits, qstep;
735         minscaler = sce->sf_idx[0];
736         //inner loop - quantize spectrum to fit into given number of bits
737         qstep = its ? 1 : 32;
738         do {
739             int prev = -1;
740             tbits = 0;
741             fflag = 0;
742             for (w = 0; w < sce->ics.num_windows; w += sce->ics.group_len[w]) {
743                 start = w*128;
744                 for (g = 0;  g < sce->ics.num_swb; g++) {
745                     const float *coefs = sce->coeffs + start;
746                     const float *scaled = s->scoefs + start;
747                     int bits = 0;
748                     int cb;
749                     float dist = 0.0f;
750
751                     if (sce->zeroes[w*16+g] || sce->sf_idx[w*16+g] >= 218) {
752                         start += sce->ics.swb_sizes[g];
753                         continue;
754                     }
755                     minscaler = FFMIN(minscaler, sce->sf_idx[w*16+g]);
756                     cb = find_min_book(find_max_val(sce->ics.group_len[w], sce->ics.swb_sizes[g], scaled), sce->sf_idx[w*16+g]);
757                     for (w2 = 0; w2 < sce->ics.group_len[w]; w2++) {
758                         int b;
759                         dist += quantize_band_cost(s, coefs + w2*128,
760                                                    scaled + w2*128,
761                                                    sce->ics.swb_sizes[g],
762                                                    sce->sf_idx[w*16+g],
763                                                    cb,
764                                                    1.0f,
765                                                    INFINITY,
766                                                    &b);
767                         bits += b;
768                     }
769                     dists[w*16+g] = dist - bits;
770                     if (prev != -1) {
771                         bits += ff_aac_scalefactor_bits[sce->sf_idx[w*16+g] - prev + SCALE_DIFF_ZERO];
772                     }
773                     tbits += bits;
774                     start += sce->ics.swb_sizes[g];
775                     prev = sce->sf_idx[w*16+g];
776                 }
777             }
778             if (tbits > destbits) {
779                 for (i = 0; i < 128; i++)
780                     if (sce->sf_idx[i] < 218 - qstep)
781                         sce->sf_idx[i] += qstep;
782             } else {
783                 for (i = 0; i < 128; i++)
784                     if (sce->sf_idx[i] > 60 - qstep)
785                         sce->sf_idx[i] -= qstep;
786             }
787             qstep >>= 1;
788             if (!qstep && tbits > destbits*1.02)
789                 qstep = 1;
790             if (sce->sf_idx[0] >= 217)
791                 break;
792         } while (qstep);
793
794         fflag = 0;
795         minscaler = av_clip(minscaler, 60, 255 - SCALE_MAX_DIFF);
796         for (w = 0; w < sce->ics.num_windows; w += sce->ics.group_len[w]) {
797             start = w*128;
798             for (g = 0; g < sce->ics.num_swb; g++) {
799                 int prevsc = sce->sf_idx[w*16+g];
800                 const float *scaled = s->scoefs + start;
801                 if (dists[w*16+g] > uplims[w*16+g] && sce->sf_idx[w*16+g] > 60)
802                     sce->sf_idx[w*16+g]--;
803                 sce->sf_idx[w*16+g] = av_clip(sce->sf_idx[w*16+g], minscaler, minscaler + SCALE_MAX_DIFF);
804                 sce->sf_idx[w*16+g] = FFMIN(sce->sf_idx[w*16+g], 219);
805                 if (sce->sf_idx[w*16+g] != prevsc)
806                     fflag = 1;
807                 sce->band_type[w*16+g] = find_min_book(find_max_val(sce->ics.group_len[w], sce->ics.swb_sizes[g], scaled), sce->sf_idx[w*16+g]);
808                 start += sce->ics.swb_sizes[g];
809             }
810         }
811         its++;
812     } while (fflag && its < 10);
813 }
814
815 static void search_for_quantizers_faac(AVCodecContext *avctx, AACEncContext *s,
816                                        SingleChannelElement *sce,
817                                        const float lambda)
818 {
819     int start = 0, i, w, w2, g;
820     float uplim[128], maxq[128];
821     int minq, maxsf;
822     float distfact = ((sce->ics.num_windows > 1) ? 85.80 : 147.84) / lambda;
823     int last = 0, lastband = 0, curband = 0;
824     float avg_energy = 0.0;
825     if (sce->ics.num_windows == 1) {
826         start = 0;
827         for (i = 0; i < 1024; i++) {
828             if (i - start >= sce->ics.swb_sizes[curband]) {
829                 start += sce->ics.swb_sizes[curband];
830                 curband++;
831             }
832             if (sce->coeffs[i]) {
833                 avg_energy += sce->coeffs[i] * sce->coeffs[i];
834                 last = i;
835                 lastband = curband;
836             }
837         }
838     } else {
839         for (w = 0; w < 8; w++) {
840             const float *coeffs = sce->coeffs + w*128;
841             start = 0;
842             for (i = 0; i < 128; i++) {
843                 if (i - start >= sce->ics.swb_sizes[curband]) {
844                     start += sce->ics.swb_sizes[curband];
845                     curband++;
846                 }
847                 if (coeffs[i]) {
848                     avg_energy += coeffs[i] * coeffs[i];
849                     last = FFMAX(last, i);
850                     lastband = FFMAX(lastband, curband);
851                 }
852             }
853         }
854     }
855     last++;
856     avg_energy /= last;
857     if (avg_energy == 0.0f) {
858         for (i = 0; i < FF_ARRAY_ELEMS(sce->sf_idx); i++)
859             sce->sf_idx[i] = SCALE_ONE_POS;
860         return;
861     }
862     for (w = 0; w < sce->ics.num_windows; w += sce->ics.group_len[w]) {
863         start = w*128;
864         for (g = 0; g < sce->ics.num_swb; g++) {
865             float *coefs   = sce->coeffs + start;
866             const int size = sce->ics.swb_sizes[g];
867             int start2 = start, end2 = start + size, peakpos = start;
868             float maxval = -1, thr = 0.0f, t;
869             maxq[w*16+g] = 0.0f;
870             if (g > lastband) {
871                 maxq[w*16+g] = 0.0f;
872                 start += size;
873                 for (w2 = 0; w2 < sce->ics.group_len[w]; w2++)
874                     memset(coefs + w2*128, 0, sizeof(coefs[0])*size);
875                 continue;
876             }
877             for (w2 = 0; w2 < sce->ics.group_len[w]; w2++) {
878                 for (i = 0; i < size; i++) {
879                     float t = coefs[w2*128+i]*coefs[w2*128+i];
880                     maxq[w*16+g] = FFMAX(maxq[w*16+g], fabsf(coefs[w2*128 + i]));
881                     thr += t;
882                     if (sce->ics.num_windows == 1 && maxval < t) {
883                         maxval  = t;
884                         peakpos = start+i;
885                     }
886                 }
887             }
888             if (sce->ics.num_windows == 1) {
889                 start2 = FFMAX(peakpos - 2, start2);
890                 end2   = FFMIN(peakpos + 3, end2);
891             } else {
892                 start2 -= start;
893                 end2   -= start;
894             }
895             start += size;
896             thr = pow(thr / (avg_energy * (end2 - start2)), 0.3 + 0.1*(lastband - g) / lastband);
897             t   = 1.0 - (1.0 * start2 / last);
898             uplim[w*16+g] = distfact / (1.4 * thr + t*t*t + 0.075);
899         }
900     }
901     memset(sce->sf_idx, 0, sizeof(sce->sf_idx));
902     abs_pow34_v(s->scoefs, sce->coeffs, 1024);
903     for (w = 0; w < sce->ics.num_windows; w += sce->ics.group_len[w]) {
904         start = w*128;
905         for (g = 0;  g < sce->ics.num_swb; g++) {
906             const float *coefs  = sce->coeffs + start;
907             const float *scaled = s->scoefs   + start;
908             const int size      = sce->ics.swb_sizes[g];
909             int scf, prev_scf, step;
910             int min_scf = -1, max_scf = 256;
911             float curdiff;
912             if (maxq[w*16+g] < 21.544) {
913                 sce->zeroes[w*16+g] = 1;
914                 start += size;
915                 continue;
916             }
917             sce->zeroes[w*16+g] = 0;
918             scf  = prev_scf = av_clip(SCALE_ONE_POS - SCALE_DIV_512 - log2(1/maxq[w*16+g])*16/3, 60, 218);
919             step = 16;
920             for (;;) {
921                 float dist = 0.0f;
922                 int quant_max;
923
924                 for (w2 = 0; w2 < sce->ics.group_len[w]; w2++) {
925                     int b;
926                     dist += quantize_band_cost(s, coefs + w2*128,
927                                                scaled + w2*128,
928                                                sce->ics.swb_sizes[g],
929                                                scf,
930                                                ESC_BT,
931                                                lambda,
932                                                INFINITY,
933                                                &b);
934                     dist -= b;
935                 }
936                 dist *= 1.0f / 512.0f / lambda;
937                 quant_max = quant(maxq[w*16+g], ff_aac_pow2sf_tab[200 - scf + SCALE_ONE_POS - SCALE_DIV_512]);
938                 if (quant_max >= 8191) { // too much, return to the previous quantizer
939                     sce->sf_idx[w*16+g] = prev_scf;
940                     break;
941                 }
942                 prev_scf = scf;
943                 curdiff = fabsf(dist - uplim[w*16+g]);
944                 if (curdiff <= 1.0f)
945                     step = 0;
946                 else
947                     step = log2(curdiff);
948                 if (dist > uplim[w*16+g])
949                     step = -step;
950                 scf += step;
951                 scf = av_clip_uint8(scf);
952                 step = scf - prev_scf;
953                 if (FFABS(step) <= 1 || (step > 0 && scf >= max_scf) || (step < 0 && scf <= min_scf)) {
954                     sce->sf_idx[w*16+g] = av_clip(scf, min_scf, max_scf);
955                     break;
956                 }
957                 if (step > 0)
958                     min_scf = prev_scf;
959                 else
960                     max_scf = prev_scf;
961             }
962             start += size;
963         }
964     }
965     minq = sce->sf_idx[0] ? sce->sf_idx[0] : INT_MAX;
966     for (i = 1; i < 128; i++) {
967         if (!sce->sf_idx[i])
968             sce->sf_idx[i] = sce->sf_idx[i-1];
969         else
970             minq = FFMIN(minq, sce->sf_idx[i]);
971     }
972     if (minq == INT_MAX)
973         minq = 0;
974     minq = FFMIN(minq, SCALE_MAX_POS);
975     maxsf = FFMIN(minq + SCALE_MAX_DIFF, SCALE_MAX_POS);
976     for (i = 126; i >= 0; i--) {
977         if (!sce->sf_idx[i])
978             sce->sf_idx[i] = sce->sf_idx[i+1];
979         sce->sf_idx[i] = av_clip(sce->sf_idx[i], minq, maxsf);
980     }
981 }
982
983 static void search_for_quantizers_fast(AVCodecContext *avctx, AACEncContext *s,
984                                        SingleChannelElement *sce,
985                                        const float lambda)
986 {
987     int start = 0, i, w, w2, g;
988     int minq = 255;
989
990     memset(sce->sf_idx, 0, sizeof(sce->sf_idx));
991     for (w = 0; w < sce->ics.num_windows; w += sce->ics.group_len[w]) {
992         start = w*128;
993         for (g = 0; g < sce->ics.num_swb; g++) {
994             for (w2 = 0; w2 < sce->ics.group_len[w]; w2++) {
995                 FFPsyBand *band = &s->psy.psy_bands[s->cur_channel*PSY_MAX_BANDS+(w+w2)*16+g];
996                 if (band->energy <= band->threshold) {
997                     sce->sf_idx[(w+w2)*16+g] = 218;
998                     sce->zeroes[(w+w2)*16+g] = 1;
999                 } else {
1000                     sce->sf_idx[(w+w2)*16+g] = av_clip(SCALE_ONE_POS - SCALE_DIV_512 + log2(band->threshold), 80, 218);
1001                     sce->zeroes[(w+w2)*16+g] = 0;
1002                 }
1003                 minq = FFMIN(minq, sce->sf_idx[(w+w2)*16+g]);
1004             }
1005         }
1006     }
1007     for (i = 0; i < 128; i++) {
1008         sce->sf_idx[i] = 140;
1009         //av_clip(sce->sf_idx[i], minq, minq + SCALE_MAX_DIFF - 1);
1010     }
1011     //set the same quantizers inside window groups
1012     for (w = 0; w < sce->ics.num_windows; w += sce->ics.group_len[w])
1013         for (g = 0;  g < sce->ics.num_swb; g++)
1014             for (w2 = 1; w2 < sce->ics.group_len[w]; w2++)
1015                 sce->sf_idx[(w+w2)*16+g] = sce->sf_idx[w*16+g];
1016 }
1017
1018 static void search_for_ms(AACEncContext *s, ChannelElement *cpe,
1019                           const float lambda)
1020 {
1021     int start = 0, i, w, w2, g;
1022     float M[128], S[128];
1023     float *L34 = s->scoefs, *R34 = s->scoefs + 128, *M34 = s->scoefs + 128*2, *S34 = s->scoefs + 128*3;
1024     SingleChannelElement *sce0 = &cpe->ch[0];
1025     SingleChannelElement *sce1 = &cpe->ch[1];
1026     if (!cpe->common_window)
1027         return;
1028     for (w = 0; w < sce0->ics.num_windows; w += sce0->ics.group_len[w]) {
1029         for (g = 0;  g < sce0->ics.num_swb; g++) {
1030             if (!cpe->ch[0].zeroes[w*16+g] && !cpe->ch[1].zeroes[w*16+g]) {
1031                 float dist1 = 0.0f, dist2 = 0.0f;
1032                 for (w2 = 0; w2 < sce0->ics.group_len[w]; w2++) {
1033                     FFPsyBand *band0 = &s->psy.psy_bands[(s->cur_channel+0)*PSY_MAX_BANDS+(w+w2)*16+g];
1034                     FFPsyBand *band1 = &s->psy.psy_bands[(s->cur_channel+1)*PSY_MAX_BANDS+(w+w2)*16+g];
1035                     float minthr = FFMIN(band0->threshold, band1->threshold);
1036                     float maxthr = FFMAX(band0->threshold, band1->threshold);
1037                     for (i = 0; i < sce0->ics.swb_sizes[g]; i++) {
1038                         M[i] = (sce0->coeffs[start+w2*128+i]
1039                               + sce1->coeffs[start+w2*128+i]) * 0.5;
1040                         S[i] =  sce0->coeffs[start+w2*128+i]
1041                               - sce1->coeffs[start+w2*128+i];
1042                     }
1043                     abs_pow34_v(L34, sce0->coeffs+start+w2*128, sce0->ics.swb_sizes[g]);
1044                     abs_pow34_v(R34, sce1->coeffs+start+w2*128, sce0->ics.swb_sizes[g]);
1045                     abs_pow34_v(M34, M,                         sce0->ics.swb_sizes[g]);
1046                     abs_pow34_v(S34, S,                         sce0->ics.swb_sizes[g]);
1047                     dist1 += quantize_band_cost(s, sce0->coeffs + start + w2*128,
1048                                                 L34,
1049                                                 sce0->ics.swb_sizes[g],
1050                                                 sce0->sf_idx[(w+w2)*16+g],
1051                                                 sce0->band_type[(w+w2)*16+g],
1052                                                 lambda / band0->threshold, INFINITY, NULL);
1053                     dist1 += quantize_band_cost(s, sce1->coeffs + start + w2*128,
1054                                                 R34,
1055                                                 sce1->ics.swb_sizes[g],
1056                                                 sce1->sf_idx[(w+w2)*16+g],
1057                                                 sce1->band_type[(w+w2)*16+g],
1058                                                 lambda / band1->threshold, INFINITY, NULL);
1059                     dist2 += quantize_band_cost(s, M,
1060                                                 M34,
1061                                                 sce0->ics.swb_sizes[g],
1062                                                 sce0->sf_idx[(w+w2)*16+g],
1063                                                 sce0->band_type[(w+w2)*16+g],
1064                                                 lambda / maxthr, INFINITY, NULL);
1065                     dist2 += quantize_band_cost(s, S,
1066                                                 S34,
1067                                                 sce1->ics.swb_sizes[g],
1068                                                 sce1->sf_idx[(w+w2)*16+g],
1069                                                 sce1->band_type[(w+w2)*16+g],
1070                                                 lambda / minthr, INFINITY, NULL);
1071                 }
1072                 cpe->ms_mask[w*16+g] = dist2 < dist1;
1073             }
1074             start += sce0->ics.swb_sizes[g];
1075         }
1076     }
1077 }
1078
1079 AACCoefficientsEncoder ff_aac_coders[] = {
1080     {
1081         search_for_quantizers_faac,
1082         encode_window_bands_info,
1083         quantize_and_encode_band,
1084         search_for_ms,
1085     },
1086     {
1087         search_for_quantizers_anmr,
1088         encode_window_bands_info,
1089         quantize_and_encode_band,
1090         search_for_ms,
1091     },
1092     {
1093         search_for_quantizers_twoloop,
1094         codebook_trellis_rate,
1095         quantize_and_encode_band,
1096         search_for_ms,
1097     },
1098     {
1099         search_for_quantizers_fast,
1100         encode_window_bands_info,
1101         quantize_and_encode_band,
1102         search_for_ms,
1103     },
1104 };